NewToolMonitor creates a new ToolMonitor.

Save stores the data to the local disk.

Applies heuristics to the SQL query.

Apply confidence thresholds from a single dict.

Assigns a value to a name (variable). Args: name: The AST node representing the name. v: The value to assign. namespace:

Generate a single loan application summary PDF.

Helper to call Model Armor REST API.

Compare two validation results to determine if violations are consistent. v2.0.2: Requires 75% overlap for confirmation.

Computes final scores and verdicts from the evaluation result table.

Convert raw shift dictionaries to Shift objects. Args: raw_shifts: List of shift dicts with keys: id, date, start, end, ward, requir

Embed an image and its 4 augmented views (left crop, right crop, flipped, text+image) in parallel.

Return empty invoice structure

Lazy-initialize GCP/Vertex AI on first use (not at import time). Agent Engine sets env vars after module import, so we must defer.

Build an error result.

Evaluates a binary operation. Args: op: The AST node representing the binary operation. left: The left operand. right: Th

Evaluate a single image (garment accuracy + wearing quality in parallel).

Extract exemption rules for a step from rules book.

Extract raw text from each slide using python-pptx.

Finds the key in a dictionary with datetime objects as keys that is closest to a given target datetime. Args: data

Fixes errors in the SQL query. Args: sql_query: The SQL query to fix. sql_dialect: The input SQL dialect. apply

INSERT staging → target, stamp timestamps, then drop staging.

Generate shift data for analysis.

Extract the agent name from the invocation context.

Get ID token headers for authenticated calls to the backend. On Cloud Run, the frontend's service account must have ``roles/run.invoker`` on

Generates a vector embedding for the given text using Vertex AI.

Get user info from access token and base64 encode the email.

Guesses the MIME type of a file based on its extension.

Lazy-initialize the Vertex AI GenerativeModel.

Retrieves the latest processed guideline output for a given guideline gcs_uri. Returns the Guideline object if found, otherwise None.

Convert seniority level to numeric order for comparison.

Lazy initialization of session folder.

Helper to get token from tool context state.

Helper function to generate a consistent log report.

Inserts a visual into a slide using a placeholder or centering it.

Check if a remote shoe classification endpoint is configured.

Checks if the object is a SQLGlot schema type.

Load embeddings and metadata into memory.

Loads an image from GCS and returns it as a Part object. Args: gcs_uri: The GCS URI of the image. Does not start with "gs://" sto

Load HRIS data.

Load nurse stats.

Write SAMPLE_RULE_BASE to a temp file and return path.

Truncate the output by omitting the middle lines. ```output line 1 line 2 (... omitted xxx lines ...) line 3

Map framing value to the model photo key prefix (e.g. 'front_top').

Return current time as HH:MM:SS.mmm.

Parse JSON from LLM response, handling markdown blocks and trailing text.

Parses a repository string to extract the owner and repository name. Handles both full GitHub URLs (http/https) and the shorthand 'owner/rep

Parse the LLM response as JSON. Handles markdown code blocks and trailing text.

Print text with word wrapping at _MAX_LINE_WIDTH.

Send one URI to the Cloud Run service, return parsed reply dict. Uses the same BQ-remote-function payload format the service already speaks (

Download and extract a single asset tar from GCS if not present locally.

Helper to run the agent and return the final response text.

Helper to run evaluation task with caching support.

Run validate → impact → auto-tighten loop on a rule. Args: case_id: The target case ID. initial_rule: The rule dict t

Run LLM validation with triple-check mechanism. 1. First call 2. If violations found + double-check enabled → second cal

Save data to JSON file.

Save data to JSON file.

Saves a JSON object as a text artifact. Args: tool_context (ToolContext): The context for saving artifacts. name (str): The base

Save nurse stats.

Rank pre-classified garment images by quality and extract details. The images have already been validated and classified by classify_garments

Try to fix an image and evaluate the fix. Returns (fixed_result_or_None, fix_log).

Validate multiple deterministic rule groups in a single LLM call. v2.0.2: Sends ALL deterministic rules in one prompt. Ultra-cons

Validate a single subjective rule group with its own LLM call. v2.0.2: Ultra-conservative prompt with triple-check.

Performs a comparative labor force assessment.

Fetches real-time news headlines to perform sentiment analysis on MSAs and industries. Use this to catch 'Soft Signals' (strikes, recent larg

Analyzes AI exposure (automation vs. augmentation) and strategic recommendations for a list of occupational domains or standard job titles.

Approve the flat rate or percentage discount requested by the user. Args: discount_type (str): The type of discount, either "percent

Build a unified case context from all agent artifacts. This flattens the multi-agent output into a single dict that the

Calculate SSIM similarity between two images after removing backgrounds. Removes white/gray backgrounds using fast HSV detection, then computes

Calls the Gemini image generation API and evaluates the result. Args: client: The Gemini API client. model: The name of the model

Call Gemini with PDF as input

Calls the Gemini model for multiple prompts in parallel using threads with retry logic. Args: prompts (List[str]): A list of prom